Prueba de validación cruzada anidada

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. Dentro de una división externa, ¿qué datos puede utilizar la búsqueda de hiperparámetros internos?
   • All examples, including the outer test fold
   • Only the outer training data
   • Only the outer test labels
   • The full dataset after fitting preprocessing globally
2. ¿Por qué informar la mejor puntuación de validación interna puede exagerar el rendimiento del modelo?
   • Inner cross-validation evaluates every candidate on its own training rows
   • Averaging across folds removes all uncertainty from candidate rankings
   • The best inner score already includes an independent outer evaluation
   • Choosing the best setting also selects favorable variation in validation scores
3. Cinco pliegues exteriores, tres pliegues interiores y diez configuraciones candidatas requieren ¿cuántos candidatos internos caben en el ejemplo de la guía?
   • 150
   • 18
   • 50
   • 30
4. Se instala un escalador en el conjunto de datos completo antes de que comience la validación cruzada anidada. ¿Qué problema persiste?
   • Using more outer folds automatically removes preprocessing leakage
   • Unsupervised preprocessing cannot use information relevant to evaluation
   • Preprocessing has already accessed held-out information
   • A scaler can safely see test inputs whenever it does not see test labels
5. Diferentes pliegues exteriores seleccionan diferentes hiperparámetros. ¿Cómo debería interpretarse esto?
   • It proves the cross-validation code is broken
   • It can occur because each selection step sees different training data
   • It means only the highest-scoring outer model is valid
   • It requires replacing outer evaluation with training accuracy
6. A dataset contains many rows per customer, and deployment targets unseen customers. Which design addresses that dependence?
   • Shuffle all rows separately in the outer loop only
   • Group customers only after evaluation is complete
   • Fit the feature selector globally before splitting
   • Keep customer groups separated in both inner and outer splits
7. After evaluating a fixed nested procedure, how does the guide recommend obtaining a final model?
   • Run the selection procedure on development data and fit the chosen configuration
   • Deploy the outer-fold model with the highest observed test score
   • Choose the configuration from the single best inner-fold result
   • Choose final settings by comparing candidates on the reserved test set
8. Why is the spread of outer-fold scores not automatically a confidence interval?
   • Each outer score is an independent measurement from an unrelated training set
   • The raw standard deviation of fold scores is itself a confidence interval
   • The fitted models share much of their training data
   • Averaging fold scores removes all sampling uncertainty